Can you steal a robot's next move by watching its clock? Journal of our experiments on timing side channels in multi agent RL
I have been keenly interested in the intersection of multi-agent reinforcement learning (MARL) and hardware security. When you deploy a trained RL policy onto a microcontroller, the model runs inference to decide what action to take. But here's the thing: different actions can take different amounts of time to compute. If an adversary can measure that timing, can they figure out what the agent is…
The article details experiments conducted on timing side channels in multi-agent reinforcement learning (MARL) policies running on microcontrollers. The core question is whether an adversary can predict an agent's action by measuring the inference duration and other timing metrics, without observing the raw observations or internal activations.
The attacker has limited visibility into the system, only being able to measure the total inference duration, per-layer inference times, network packet timings (if WiFi is used), and not the raw observations or model weights. The threat is significant because MARL is increasingly deployed on edge devices, potentially giving adversaries physical proximity to tap into timing signals.
The article presents findings from three test environments: Cooperative Grid Navigation, CartPole, and MPE Simple Spread. The grid navigation scenario shows the most leakage, with Mutual Information (MI) values around 1 bit per action, indicating that an attacker could predict actions with high accuracy using timing alone. The classifiers trained on timing features achieved near-perfect accuracy, confirming that the timing leakage can be exploited for malicious purposes.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — it may contain errors, so check the original before relying on it.
